New model.yaml activation: field

[glenn-jocher]

Add optional model yaml activation field to define model-wide activations, i.e.:

activation: nn.LeakyReLU(0.1)  # activation with arguments
activation: nn.SiLU()  # activation with no arguments

Signed-off-by: Glenn Jocher glenn.jocher@ultralytics.com

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced flexibility in activation functions and depth-wise convolutions in YOLOv5 models.

📊 Key Changes

• Introduced a preset default activation function (SiLU) for the Conv class.
• Modified the DWConv class to support dilation, making it more versatile.
• Added a new YOLOv5s configuration file utilizing the LeakyReLU activation function.
• Implemented the ability to set a custom activation function globally for YOLOv5 models through the model configuration YAML.

🎯 Purpose & Impact

• 🎛 The default SiLU activation simplifies the construction of Conv layers by establishing a consistently used non-linearity.
• ⚙️ Depth-wise convolutions now have the added flexibility of specifying dilation, thereby enhancing the ability to control feature map resolution during the learning process.
• 🔌 The introduction of a new YOLOv5s model using LeakyReLU provides users with a pre-configured alternative if it better suits their use case.
• 📐 With the update, developers can easily swap activation functions across the entire model, promoting experimentation and potentially leading to performance improvements.
• 👩‍💻 Users benefit from increased model customization abilities, potentially leading to better accuracy, training speed, or inference efficiency, depending on the task at hand.

[glenn-jocher]

Refer to ultralytics/hub#97

[glenn-jocher]

@AyushExel tracking YOLOv5m DetectionModel activation study in https://wandb.ai/glenn-jocher/YOLOv5m-study-activations

Note yolov5m-SiLU is lagging on CUDA device 5 unfortunately, runs about half as fast as other devices.